______ on Display is an exploration of entropy in the SD v1.5 model.

What happens when you use high-entropy input images (generated on the Pico-8 Fantasy Console), with high-entropy prompts (with blanks and incomplete phrases)...

Despite different prompts or inputs, the model will often resort to some common motifs that appear again and again. Many of these are haunting images or figures. I like to think that this is a way to give the model a Rorschach test. You're not only seeing what the model "sees" in the input image... but also what it has seen and learned about, and stored in its own weights.

Because of the high entropy, there are also failure states, which I find just as interesting as the more "robust" looking outputs. Why does the model fail on some occasions but succeed on others? What difference do just a few pixels or colors in the input image make on the process the AI uses to generate its images?

Finally, when you see these displays, and you see what the AI is generating, you yourself have to interpret what you're seeing. Is the model really haunted? Or is it the human who brings their inner ghosts to the interpretations?